V

1
Víctor Ponce Miguel Reyes Xavier Baró Mario
Gorga Sergio Escalera
Two-level GMM Clustering of Human Poses for
Automatic Human Behavior Analysis
Abstract
Detect human poses is a main step in the study of
human behavior analysis. Our achievement is to
find a non supervised method to determine key
poses of a human gesture for a posterior analysis
of a human behavior using machine learning and
computer vision techniques. We use Kinect system
for body feature extraction from joint positions
of the articulated human models and then we apply
a global and local feature clustering using a
two-level Gaussian Mixture Model approach for the
expert evaluation by comparing each clustering
levels through their visualizations. Keywords
Human Pose, Human Behavior Analysis, Clustering.
Methodology
Human Pose Representation
Human Pose Clustering

Grouping the previous pose descriptions in pose
clusters with standard Gaussian Mixture Model.
Our goal is to group the set of frame pose
descriptions in clusters so that posterior
learning algorithms can improve generalization in
Human Behavior Analysis systems. We use a full
covariance GMM of K components parameterized
Depth maps acquired by public API OpenNI software
1. These features are able to detect and track
people to a maximum distance of six meters from
multi-sensor device. Method of 2 used for the
detection of human body and its skeletal model.
The approach of 2 uses a huge set of human
samples to infer pixel labels through Random
Forest estimation, and skeletal model is defined
as the centroid of mass of the different dense
regions using mean shift algorithm. The
articulated human model is defined by the set of
15 reference points. This model has the advantage
of being highly deformable, and thus, able to fit
to complex human poses.
Then, a likelihood value based on the probability
distributions p() of the GMM is obtained

• First level Clustering Use three spatial
  components of descriptor V for each joint i, i ?
  1, , 14 and perform GMM of k1 clusters,
  namely GMM i1
• Second level Clustering
• Define for each pose a new feature vector,
  of size
  14xk1, where vji is the probability of applying
  GMM model GMMi1 at features from V corresponding
  to spatial coordinates of j.th joint.
• Use components of descriptor V and perform GMM of
  k2 clusters, namely GMM2.

 
Data Set of gestures using the Kinect device
consisting of seven different categories. It has
been considered 10 different actors and different
environments, having a total of 130 data
sequences with 32 frame gestures. Thus, the data
set contains the high variability from
uncontrolled environments. The resolution of the
video depth sequences is 340x280. Methods and
parameters The people detection system used is
provided by the public library OpenNI. This
library has a high accuracy in people detection,
allowing multiple detection even in cases of
partial occlusions. Figure 3 shows consecutive
visual descriptions of some data set gestures. At
the bottom of the sequences, we show a first row
that represents the cluster number assigned by a
one-level GMM, and a second row with the assigned
cluster using two-level GMM. One can see that in
most cases both grouping techniques assigns
consecutive poses to same clusters. However, the
clusters assigned by the one-level GMM have more
visual variability, being inefficient for human
behavior generalization purposes.
Departament de Matemàtica Aplicada i Anàlisi,
Universitat de Barcelona, Spain Centre de Visió
per Computador, Universitat Autònoma de
Barcelona, Spain E-mail v88ponce_at_gmail.com,
mreyese_at_gmail.com, xbaro_at_cvc.uab.es,
mario.gorga_at_gmail.com, sergio_at_maia.ub.es